How to review a poker decision without grading the outcome

Record the decision before the result can rewrite it

A useful review asks whether an action followed from the information and constraints available at the time. A win is one realized outcome; it does not reveal the quality of the probability model by itself. Equally, a loss does not establish that the chosen action was a mistake.

Before revealing an answer or observing a simulated outcome, write down the pot definition, incremental cost, possible payoffs, estimated probabilities and alternatives. Mark which quantities were supplied by the exercise and which you inferred. Keeping those categories separate prevents a later result from silently becoming an input to the original decision.

Use this as a learning log, not a claim about real trading performance. The examples below use invented chip payoffs and a fully specified probability. Real estimates may be wrong, observations may be dependent and the set of available actions may be incomplete.

Work one choice from net cash flows

Consider a final call costing 40 into a pot of 120, with a given 30% chance of winning. There are no ties, fees or later bets. Relative to folding, winning gives +120 and losing gives −40. The expected result is +8 chips: 0.30×120−0.70×40. Returning your own call after a win is not additional profit.

The 25% break-even threshold is a price calculation. A 30% belief clears that threshold under this model, but a 20% belief would not. If you later discover that an out was not actually clean, revise the model rather than retaining the original probability and blaming variance.

Checkpoint: verify the calculation at probability zero and one. You should get −40 and +120 respectively. A formula that produces different endpoints has probably mixed gross pot size with net payoff.

E[X] = 0.30(120) + 0.70(−40) = 8

An advantage can coexist with a losing short sample

Repeat this exact toy decision ten times, independently, without changing its win probability. Let K count wins. Total chips are 120K−40(10−K)=160K−400. Two or fewer wins produce a loss; three wins produce +80. The expected total is +80, but that is not the only plausible result.

Enumerating the cases K=0,1,2 gives a probability of about 38.28% for a losing ten-decision sample. That is compatible with positive expectation. The number is specific to this experiment; it is not a universal loss rate for poker or a forecast for a learner.

The independence and fixed-probability assumptions matter. Repeatedly facing the same uncertain model, changing the stakes or selecting only memorable hands can invalidate this calculation. A streak is not by itself evidence for independence or for a stable win chance.

P(K≤2) = 0.7¹⁰ + 10(0.3)0.7⁹ + 45(0.3²)0.7⁸ ≈ 0.382783

Measure dispersion instead of calling it bad luck

For one repetition, the squared-payoff average is 0.30×120²+0.70×40²=5,440. Subtracting the squared mean, 8², gives variance 5,376 chips squared. The standard deviation is about 73.32 chips, much larger than the eight-chip mean.

For 100 independent repetitions with that same distribution, the standard deviation of the sample mean is about 7.33 chips. This describes variability under the assumed model. It is not a confidence guarantee about an unknown real-world edge, and it does not include model error.

Do not use a larger sample as a reason to stop examining assumptions. An estimate can become precisely wrong if observations are biased or the environment changes. Compare exact calculations, simulation output and the data-generating story before drawing a conclusion.

Var(X) = E[X²] − E[X]² = 5440 − 64 = 5376

Use a short, auditable review template

First, copy the information set: what was visible and what was hidden? Second, state the alternatives and their net payoffs. Third, record the probability model and its source. Fourth, calculate expectation and a sensitivity case. Fifth, write the action and any binding constraint. Only then record the realized result.

Afterward, classify a disagreement. Was it a rules error, a counting error, a payoff-definition error, an unsupported probability estimate or a reasonable decision followed by an unfavorable realization? These categories call for different practice. Repeating arithmetic does not fix a mistaken information model.

Keep a correction next to the original record rather than replacing the original with what you wish you had believed. In a hosted quiz, use the explanation after reveal to identify the mistaken step; the scoreboard alone cannot tell you why an answer was wrong.

Choose the next exercise from the failure, not the score

If the issue was a denominator, return to deck counting. If you conflated a price with a belief, derive the final-call threshold again. If you updated an opponent model after too few observations, examine base rates and conditional likelihoods. If you cannot state what would falsify your claim, design a small counterexample.

Account lesson progress on Poker4Quant summarizes your latest explicitly saved checks, not an immutable decision log or a real-money performance record. Keep a separate explanation in your own notes when you want to study how a belief evolved. Device guest answers and account snapshots also remain separate until you explicitly save.

A useful stopping point is being able to explain a changed version of the problem without looking at its answer. That is a practical review criterion, not a credential. Revisit the calculation when the sampling rule, payoffs or information set changes.

Sources and further reading

• MIT OpenCourseWare — Probability and Statistics readings ↗

  Background on expectation, variance, binomial models and sample averages. This article’s chip experiment, log template and interpretation are original.